Research On Object Tracking By Kernel Correlation Filter Based On HCF

Object tracking is widely applied in the field of life,traffic and medical treatment.However,it is challenging as that objects often undergo significant appearance changes caused by deformation,abrupt motion,background clutter and occlusion.In recent years,the application of deep learning to object tracking algorithm has become popular.In this paper,combining convolution neural network with kernel correlation filter as a basis.This paper makes an in-depth analysis and research from three aspects: improving the real-time performance,scale adaptive and occlusion.The main research work of this thesis is as follows:Firstly,an improved method for multi-layer feature extraction of convolutional neural networks is proposed.The number of filters in the convolutional layer determines the number of feature channels extracted.Making structural adjustment by reducing the number of filters in each convolutional layer.Using the pre-trained network for feature extraction,and applying the single layer feature to the kernel correlation filter tracking algorithm.After in-depth analysis of each layer feature by tracking result comparison and feature visualization,the algorithm uses three-layer convolution network to extract the feature and three correlation filter are trained respectively.Finally,the tracking results are weighted to achieve the accurate position of the target.Secondly,a scale adaptive target tracking method based on hierarchical convolution features is proposed.It divides the object tracking into two parts: position estimation and scale estimation.The former uses the multi-level convolution features combined with the correlation filter algorithm to locate the target accurately.The scale estimation is based on the target position estimated by the convolution feature.The Edge Box algorithm is used to detect the candidate proposals.Candidate box scores are calculated based on the contours completely contained in the candidate bounding box.Seveal candidate boxes with higher scores are retained and convoluted with the correlation filter.The size of the candidate proposals with the largest response value is the size of the bounding box.Thirdly,a long-term adaptive target tracking method based on hierarchical convolution features is proposed.In the KCF framework,the peak sidelobe ratio is used as the tracking confidence.When the peak sidelobe ratio is greater than the set threshold,the tracking result is reliable enough,the scale adaptive algorithm is used to estimate the new target size,and the model is adaptively updated;When the peak sidelobe ratio is less than the set threshold,the tracking result is inaccurate due to problems such as occlusion and rotation.At this time,the target is re-detected by the random fern classifier,so as to solve the problem of target drift and target loss that often occur in the process of long-term object tracking.In this paper,part of the public video sequence on OTB is selected for simulation of the above algorithm,and comparison with other popular algorithms is made for qualitative analysis and quantitative analysis.The results show that improved convolution neural network feature extraction,the speed and accuracy of tracking can be effectively improved;The scale adaptive object tracking based on hierarchical convolution features can effectively solve the problem of scale change of targets;The long-term object tracking based on hierarchical convolution features can effectively solve the problems of target drift and target loss in the process of long-term target tracking by re-detecting the target.